Into the Impossible With Brian Keating

Meta NeuroScientist SHOCKED Me: Scale Alone Won’t Create Consciousness!

50 snips

Mar 16, 2026

David Sussillo, Meta Reality Labs research scientist and Stanford adjunct who built RNN training methods and brain-computer interfaces. He argues that scaling transformers alone will not produce consciousness. He contrasts RNNs and transformers, explains Meta’s EMG wristband work, and discusses the FORCE algorithm, lock-in to GPU/LLM paradigms, and why brain-inspired approaches may be needed.

Ask episode

AI Snips

Chapters

Books

Transcript

Episode notes

GPU Fit Fueled Transformer Dominance

The transformer architecture was designed to run efficiently on GPUs by using attention and large parallel matrix multiplies.
That hardware–architecture fit accelerated transformer dominance across ML applications.

LLM Plus GPU May Create Lock-In

Brian Keating argues 'lock-in' from LLM+GPU marriage may cap AI's ceiling and prevent alternative architectures from emerging.
He uses historical lock-in examples like QWERTY and VHS to illustrate path dependence.

Validate Scaled Models On Diverse Biological Data

Use deep learning at scale to decode biological signals but test generality across people.
Sussillo's EMG wristband work showed log-linear scaling improvements and published results in Nature validating DL on muscle signals.

Get the Snipd Podcast app to discover more snips from this episode